3D Target Structure Tracking from 2D Projections in Radiotherapy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for tracking planning target volumes (PTVs) and organs at risk (OARs) in radiotherapy are limited by manual template extraction, human error, and inability to accurately estimate motion in low-contrast regions, leading to uncertainty and potential damage to healthy tissues.
Innovation Solution
An end-to-end target structure tracking system using machine learning models to automatically track PTVs and OARs in three-dimensional space, leveraging dual domain conversion of 2D projection data to 3D space, incorporating reconstruction algorithms and recurrent neural networks to predict probability distributions of their locations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional manual template extraction and tracking methods are used, then the process is simple and straightforward, but human error and bias result in large uncertainty margins in PTV and OAR tracking
Solution Approach 1:
The patent replaces manual mechanical/template-based tracking methods with an end-to-end deep learning system that automatically extracts features and tracks PTVs and OARs. The neural network models substitute the manual template extraction process, eliminating human error and bias while achieving superior tracking accuracy through automated feature learning from imaging data.
Solution Approach 2:
The system performs self-service by automatically extracting templates and tracking structures without requiring manual intervention. The deep learning models autonomously learn optimal feature representations and perform tracking, making the system self-sufficient and eliminating the need for operator skill-dependent manual template creation.
2Measurement precision
If conventional 2D tracking methods are used, then the implementation is straightforward, but the system is limited in locating PTV and OAR in three-dimensional space
Solution Approach 1:
The patent transitions from 2D projection-based tracking to 3D volumetric tracking by implementing an end-to-end deep learning system that operates in three-dimensional space. The model processes 3D imaging data and outputs 3D localization coordinates for PTVs and OARs, enabling accurate spatial tracking that captures the full anatomical context rather than limited 2D projections.
3Measurement precision
If conventional reconstruction algorithms are used, then the processing is faster, but they cannot accurately estimate the motion of PTV and OARs in low-contrast regions
Solution Approach 1:
The patent replaces conventional reconstruction algorithms with deep learning-based motion estimation models. The neural networks automatically learn to estimate PTV and OAR motion from imaging sequences, achieving superior accuracy in low-contrast regions by learning robust feature representations that are insensitive to contrast variations, while maintaining efficient processing speeds through optimized network inference.
4Reliability
If manual template extraction is performed, then the user can control the area of interest, but human error and bias create large uncertainty margins in tracking
Solution Approach 1:
The system performs self-service by automatically extracting templates and tracking structures without requiring manual intervention. The deep learning models autonomously learn optimal feature representations and perform tracking, making the system self-sufficient and eliminating the need for operator skill-dependent manual template creation, thereby improving reliability while maintaining ease of operation.
Solution Approach 2:
The patent replaces manual mechanical/template-based tracking methods with an end-to-end deep learning system that automatically extracts features and tracks PTVs and OARs. The neural network models substitute the manual template extraction process, eliminating human error and bias while achieving superior tracking accuracy through automated feature learning from imaging data.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments described herein provide for determining a probability distribution of a three- dimensional point in a template feature map matching a three-dimensional point in space. A dual-domain target structure tracking end-to-end system receives projection data in one dimension or two dimensions and a three-dimensional simulation image. The end-to-end system extracts a template feature map from the simulation image using segmentation. The end-to-end system extracts (206) features from the projection data, transforms the features of the projection data into three-dimensional space, and sequences the three-dimensional space to generate a three-dimensional feature map. The end-to-end system compares (210) the template feature map to the generated three-dimensional feature map, determining (212) an instantaneous probability distribution of the template feature map occurring in the three-dimensional feature map.